search

happynear / AMSoftmax

A simple yet effective loss function for face verification.
MIT License
485 stars 130 forks source link

high similarity between abnormal face images #32

Open jyshee opened 5 years ago

jyshee commented 5 years ago

Hi, I found that the model trained with AMS may get higher similarity bettem a pair of abnormal enroll and probe images (the probe is low qulaity, wrong aligned, or even not a face, the enroll is not a good id photo). The similarity may be around 0.4 or even higher while the ones trained with softmax may be just around 0. So have you ever met the same problems? Is it because that the margin push the feature space much compact than softmax? Thanks!

happynear commented 5 years ago

AMS indeed push the scores towards 1.0. However, it should not be so big (0.4). I think this is a special case.

You may check the threshold calculated by the evaluation codes. The threshold reveals how much the scores shift.

jyshee commented 5 years ago

Do you mean the threshold of a certain FPR? Then the threshold is high because of such abnormal fp imgs.